Skip to content
ConsultEvo

Microsoft Power Automate: Flow Types, Uses, and Limits

Microsoft Power Automate is Microsoft’s workflow and process-automation service within the Power Platform. It can respond to events, run on a schedule, start from a user action, or automate tasks in supported Windows applications. It is worth considering when a process has clear triggers, actions, and decision rules, particularly when Microsoft 365, SharePoint, Teams, Dataverse, Power Apps, or Windows desktop applications are central to the work.

The practical answer is not simply to automate the first sequence of clicks. First map the process, identify its owner and system of record, define the desired outcome, and decide what happens when the normal path fails. Then verify the required connector, operation, authentication method, license, environment, machine, and retention controls.

A business process flow is a separate feature. It guides people through stages and steps in a model-driven Power Apps application. It is not interchangeable with a background cloud flow that reacts to an event.

Automate a defined process with an owner, a source of truth, and explicit exception paths, not merely a sequence of clicks.

Choose the right Power Automate flow type

Microsoft’s flow-types overview describes cloud flows, desktop flows, and generative actions as principal categories. Business process flows are documented separately because they guide users through stages in model-driven applications.

Capability How it starts or works Use it when
Automated cloud flow A supported event triggers cloud actions. A new record, message, or other event should start the process.
Instant cloud flow A user starts it, such as with a button. A person should choose when to run the action.
Scheduled cloud flow It runs at a configured time or interval. A recurring check, report, or synchronization is required.
Desktop flow It automates a user interface on a supported Windows machine. The application lacks a suitable connector or interface.

Business process flows provide stage and step guidance in model-driven Power Apps. Required steps can act as stage gates. Choose one when people need a visible, consistent path through a process, not when a background event should update records without user interaction. See Microsoft’s business process flow overview.

Process mining analyzes event-log data from systems of record. Task mining examines desktop work and user actions. Both can reveal paths, variations, and bottlenecks, but their findings are inputs to a separate automation design. They do not automatically produce a safe production flow. Microsoft documents these capabilities in its process mining overview and task mining overview.

A practical workflow: build an approval around a source record

A durable approval starts with a source record, not only an email or approval card. Microsoft’s SharePoint approval tutorial documents a sequence in which a SharePoint item is created, an approval is started and awaited, the outcome is evaluated, the requester is notified, and the originating item is updated.

For an illustrative purchase request, the source item might contain RequestId, Requester, Approver, Amount, Status, ApprovalCycle, Decision, and Comments. These are proposed field names, not Microsoft-required columns. Validate required values and the approver identity before sending the approval.

01Create the requestWrite the request fields, stable source ID, requester, approver, and pending status to the SharePoint item.
02Start and wait for approvalSend the approval to the resolved approver and retain the response, comments, and approval-cycle identifier.
03Evaluate the outcomeBranch on the actual response, including rejection, cancellation, timeout, or an unresolved approver.
04Update and notifyWrite the decision and comments to the source item, update its status, and notify the requester.
05Escalate exceptionsThe process owner handles missing approvers, cancellations, timeouts, deleted items, and approvals that remain unresolved.

If a request can be resubmitted or routed to several approvers, separate the request record from approval cycles and individual responses. A proposed key could combine the source record ID with an approval-cycle number. If concurrent runs or retries can create duplicates, use a destination-enforced unique constraint or transactional upsert. A lookup followed by create is not race-safe.

Microsoft advises storing approvals expected to last more than 30 days in Dataverse. Treat that as a design requirement to check when an approval may remain open for an extended period, rather than assuming a short-lived approval pattern covers every case.

Use AI Builder for extraction, not as the approval decision

Microsoft’s AI Builder document-processing example uses the Process documents action with a trained model. The flow can use extracted fields, table values, and confidence scores in later actions. For an illustrative invoice, the model might return candidate values such as these:

{
  "InvoiceNumber": "INV-1042",
  "SupplierId": "SUP-18",
  "InvoiceDate": "2026-11-22",
  "Currency": "USD",
  "Total": 1250,
  "FieldConfidence": {
    "InvoiceNumber": 0.98,
    "Total": 0.76
  }
}

The values and field structure above are illustrative. Use AI Builder to propose extracted values, then use deterministic rules to decide whether they are acceptable. Require an invoice number and supplier, check the supplier against an approved list, validate date and currency formats, and compare the total with line items where available.

Extraction is not validation

A confidence score can route a record for review, but it does not prove that the value is correct. Send missing, contradictory, or low-confidence fields to a named accounts-payable reviewer before writing an accepted invoice to the system of record.

A practical intake sequence is: receive the file, pass its content to the trained model, validate candidate fields, write only validated data to the destination, and route exceptions to review. Preserve the source-file reference, content fingerprint, model name and version, processing time, extracted values, confidence values, validation status, and reviewer decision. A fingerprint can help identify repeated submissions, but it is a proposed control rather than an automatic duplicate check supplied by the example.

Microsoft documents PDF, JPG, and PNG inputs with a maximum document size of 20 MB. Other requirements include language, page, and field limitations. Review the current AI Builder document-processing requirements before accepting files or setting a model scope.

Desktop RPA is useful for UI work, but constrained by the machine

A cloud flow can invoke a desktop flow through Microsoft’s documented cloud-to-desktop flow pattern. The general sequence is to configure the desktop flow, register a physical or virtual machine or machine group, configure a connection, select attended or unattended execution, invoke the flow from a cloud flow, and monitor the result.

For an illustrative legacy case-entry task, a cloud event could pass a case ID to a desktop flow. The desktop flow would enter the data into the target application and return a status or application reference. The source system should store the run ID, machine or machine-group reference, desktop-flow version, result, and exception path. The event source, target application, output fields, and destination must be verified for the deployment.

Prefer robust selectors or an application interface over fixed mouse coordinates where possible. Test the intended screen resolution and distinguish a transient connection problem from a deterministic application failure. A changed screen, missing selector, expired credential, or application timeout needs an operator or application owner to investigate rather than an automatic business retry.

Unattended execution is not a background mode available on any PC. It requires suitable licensing, an available registered machine, credentials, permissions to create the required remote desktop session, and compatible Windows configuration. Microsoft states that Windows 10 and Windows 11 devices cannot run unattended desktop flows while active Windows user sessions remain, including locked sessions. Windows Home editions can author and locally run attended flows but cannot trigger desktop flows from the cloud. Review Microsoft’s unattended desktop flow requirements and Power Automate for desktop system requirements for the target environment.

Licensing: check the exact use right before estimating cost

Selected Microsoft 365 or Office 365 licenses provide use rights for some Power Automate scenarios using standard connectors. That does not mean every subscription includes premium or custom connectors, attended or unattended desktop orchestration, or every cloud-flow capability. Check the exact subscription and operation required by the process.

As of November 22, 2026, Microsoft’s U.S. pricing page lists Premium at $15 per user per month, Process at $150 per bot per month, and Hosted Process at $215 per bot per month when paid yearly. Premium covers relevant premium and custom connector use and attended desktop flows. Process includes unattended RPA. Hosted Process adds a Microsoft-hosted virtual machine. Microsoft states that prices vary by country, currency, region, and organizational variant. The pricing page advertises a 30-day trial, subject to availability and eligibility.

These are current U.S. list prices, not a universal quote or total project cost. Include environment administration, machine infrastructure, testing, monitoring, connector maintenance, security review, and specialist support in the estimate. Consult Microsoft’s license entitlement comparison and licensing FAQ before purchase.

Verify before choosing a plan
  • Does the required connector and operation use standard, premium, or custom connector rights?
  • Will the flow start from an event, a user action, or a schedule?
  • Does desktop work require attended or unattended execution?
  • For unattended work, is a supported machine available, and can active user sessions be excluded?
  • Do the current regional price, tenant terms, action limits, and organizational requirements match the estimate?

Where Power Automate fits and what to compare

Power Automate is a strong candidate when Microsoft 365, SharePoint, Teams, Dataverse, model-driven Power Apps, or Windows desktop automation are material requirements. In a mixed-application environment, choose based on the specific job rather than a universal claim that one platform is best.

Compare the exact trigger and action, authentication, writable fields, environment availability, pagination, throttling, update behavior, and licensing. A product page may show that a service is supported while the required operation, account type, or write capability still needs testing.

For teams evaluating cross-application automation, ConsultEvo’s Zapier automation service is one option to review alongside the requirements. Compare shortlisted approaches by the real source and destination, exception handling, access controls, operational ownership, and ongoing cost. Do not rely on fixed connector counts or broad claims about which platform is easier or best.

Run a proof of concept with controlled representative data. Test creation and update behavior, authentication renewal, throttling, a failed action, a retry, and the destination’s handling of duplicate submissions. Do not promise a direct integration until the exact connector or API operation is confirmed.

Production controls and measurable outcomes

Before release, name the process owner, system of record, exception queue, retry behavior, and operating contact. Record stable source IDs alongside useful run details such as flow version, run ID, source reference, result, and exception status. Check permissions, privacy, and retention before processing request data, documents, event logs, or desktop recordings.

Make writes idempotent when a flow can retry. The same logical input should not create a second business record. Where concurrent runs are possible, enforce uniqueness at the destination or use a transactional upsert. For document intake, a source-system and content fingerprint can identify repeat files. For approvals, a source request ID plus cycle number can distinguish resubmissions, while a separate response record can preserve each approver’s action.

Keep data at the correct grain. An approval response is an event for one approver and one cycle. A document record is a validated business object. A daily exception-rate report is an aggregate and should not be stored as though it were an event. For process mining, use a consistent event structure such as case ID, activity, and timestamp, and keep summary metrics in a separate aggregation.

Measure outcomes that match the process. Proposed measures include elapsed time to decision, successful completion rate, exception rate, duplicate-write rate, and manual review volume. These are implementation measures, not vendor-reported results. Use them to decide whether the process is improving, not merely whether a flow ran.

Bottom line

Use Power Automate when the process can be expressed as clear triggers, actions, and decision rules, and when its connectors, licensing, and runtime conditions fit the work. Cloud flows handle connector-based event, user-started, and scheduled tasks. Desktop flows automate supported Windows UI work. Business process flows guide users through stages in model-driven apps.

A reliable rollout starts with a mapped process, a source of truth, explicit validation, stable identifiers, and a named owner for exceptions. Treat AI Builder as an extraction aid, desktop RPA as machine-dependent automation, and licensing as a design input to verify before committing.